GitHub - vicariousinc/science_rcn: Reference implementation of a two-level RCN m...
source link: https://github.com/vicariousinc/science_rcn
Go to the source link to view the article. You can view the picture content, updated content and better typesetting reading experience. If the link is broken, please click the button below to view the snapshot at that time.
Reference implementation of Recursive Cortical Network (RCN)
Reference implementation of a two-level RCN model on MNIST classification. See the Science article "A generative vision model that trains with high data efficiency and breaks text-based CAPTCHAs" and Vicarious Blog for details.
Note: this is an unoptimized reference implementation and is not intended for production.
Note: Python 3.9 is supported. The code was tested on OSX 12.3.1. It may work on other system platforms but not guaranteed. You will need the packages listed in
requirements.txt to be installed.
Clone the repository:
git clone https://github.com/vicariousinc/science_rcn.git
The code is pure Python, so you can run it right away, although you will have to uncompress the ZIP in the data folder manually.
Alternatively, install with (setting up a virtual environment beforehand is recommended):
python setup.py install
If you installed via
make you need to activate the virtual environment:
To run a small unit test that trains and tests on 20 MNIST images using one CPU (takes ~2 minutes, accuracy is ~60%):
To run a slightly more interesting experiment that trains on 100 images and tests on 20 MNIST images using multiple CPUs (takes <1 min using 7 CPUs, accuracy is ~90%):
python science_rcn/run.py --train_size 100 --test_size 20 --parallel
To test on the full 10k MNIST test set, training on 1000 examples (could take hours depending on the number of available CPUs, average accuracy is ~97.7+%):
python science_rcn/run.py --full_test_set --train_size 1000 --parallel --pool_shape 25 --perturb_factor 2.0
Check out our related blog post.
We used the following datasets for the Science paper:
Yann LeCun (Courant Institute, NYU) and Corinna Cortes (Google Labs, New York) hold the copyright of MNIST dataset, which is a derivative work from original NIST datasets. MNIST dataset is made available under the terms of the Creative Commons Attribution-Share Alike 3.0 license.
Aggregate valuable and interesting links.
Joyk means Joy of geeK